Matched filter estimation of serial blood vessel diameters from video images

Abstract

A method for making a contiguous series of blood vessel diameter estimates from digitized images is proposed. It makes use of a vessel intensity profile model based on the vessel geometry and the physics of the imaging process, providing estimates of far greater accuracy than previously obtained. A variety of techniques are used to reduce the computational demand. The method includes the generation of measurement estimation error, which is important in determining total vessel patency as well as providing a basic measure of diameter estimate accuracy.

Retinal Vessel Centerline Extraction Using Multiscale Matched Filters, Confidence and Edge Measures

“…Many published algorithms for optic disc detection [12,19,65], image registration [5,41,45,52,59,69,71], change detection [16,18,21,22,50,58,70], pathology detection and quantification [25], tracking in video sequences [33,43,55], and computer-aided screening systems [29,46,60,68] depend on vessel extraction. The techniques published in the research literature in response to the importance of retinal vessel extraction may be roughly categorized into methods based on matched filters [8,20,43], adaptive thresholds [20,24], intensity edges [4,32], region growing [40], statistical inferencing [38], mathematical morphology [18,40,72], and Hessian measures [3,25,56,57]. This wide range of techniques closely corresponds to the suite of methods that have been applied throughout the medical image analysis literature [28].…”

A piecewise Gaussian model for profiling and differentiating retinal vessels

“…As edge fitting is less sensitive to noises, it is employed in vessel detection in many research works. Negative step gate function [5], blurred half-ellipse [8], and Gaussian function [3] [4] [9] have been proposed as models for vessel profiling. Among them, the single Gaussian function is the most popular model.…”

Fourier cross-sectional profile for vessel detection on retinal images

“…Kernel-based methods work by convolving images with a filter kernel defined by the model of vessel cross-sectional profile. Blurred half-elliptical profile [14], Gaussian shaped profile [7,15,16], and simple rectangular profile [17] have been proposed for modeling profile cross-sections. A popular detection methodology is to construct optimal matched filters with the shape of a Gaussian profile [7,16,18].…”